I-Projection Results

Steven

2025-09-30

Overview

Under the hood

What this analysis does

  • Loads a RoBERTa (a LLM) to embed participant text into high-dimensional vectors.

  • Defines psychological axes (valence, ability) using controlled “I am” / “I feel” statements; each axis is the normalized difference between mean embeddings of positive vs. negative statements.

  • This allows us to capture the psychological dimensions of interest in the embedding space (i.e., good-bad and able-unable).

  • Projects target words (e.g., I, me, my) onto these axes by computing the dot product between each word’s embedding and the axis vector, capturing how strongly that word aligns with a psychological dimension.

  • Aggregates these projections at the text level to create text-level measures of valence and ability projections for each participant, which we can then use for down-stream analyses.

Q1.) Do projections move closer to the positive end of the axis over time?

Note. Data are aggregated at the assessment level and coefficients are unstandardized.First person sing. = means of I, me, and my projections for each participant.

first_person_sing_valence_combined

first_person_sing_valence_combined individual variation

first_person_sing_valence_combined spline fit

First-person sing. able projections (am, feel, & combined)

Note. Data are aggregated at the assessment level and coefficients are unstandardized.First person sing. = means of I, me, and my projections for each participant.

first_person_sing_ability_combined

first_person_sing_ability_combined individual variation

first_person_sing_ability_combined spline fit

Q2.) Do our projection measures predict changes in internalizing symptoms over time?

Results are condensed to be for the first-person singular valence and ability projections only (combined).

Valence Projection

Ability Projection

Q3.) Are these predictions robust to other linguisitic features?

Results are condensed to be for the first-person singular valence and ability projections only (combined).

Valence Projections

Ability Projections

Q4.) Can we cluster based on projection scores?

Good-bad

Map Change over time in each cluster

# Standardization method: pseudo

Parameter          | Std. Coef. |       95% CI
----------------------------------------------
(Intercept)        |       0.00 | [0.00, 0.00]
daysSinceFirstText |       0.02 | [0.02, 0.03]
# Standardization method: pseudo

Parameter          | Std. Coef. |         95% CI
------------------------------------------------
(Intercept)        |       0.00 | [ 0.00,  0.00]
daysSinceFirstText |      -0.02 | [-0.03, -0.01]

Differences in Symptoms Across Clusters

Controlling for baseline symptoms

Plot

Do clusters moderate?

model = lmer(Internalizing ~ first_person_sing_valence_combined * clusterComb + daysSinceFirstText + (1|room_id), data = data_cluster)

Table

Plot

All plots together

Able-Unable

Map Change over time in each cluster

# Standardization method: pseudo

Parameter          | Std. Coef. |       95% CI
----------------------------------------------
(Intercept)        |       0.00 | [0.00, 0.00]
daysSinceFirstText |   9.98e-03 | [0.00, 0.02]
# Standardization method: pseudo

Parameter          | Std. Coef. |       95% CI
----------------------------------------------
(Intercept)        |       0.00 | [0.00, 0.00]
daysSinceFirstText |       0.05 | [0.04, 0.06]
# Standardization method: pseudo

Parameter          | Std. Coef. |       95% CI
----------------------------------------------
(Intercept)        |       0.00 | [0.00, 0.00]
daysSinceFirstText |       0.08 | [0.06, 0.10]
# Standardization method: pseudo

Parameter          | Std. Coef. |       95% CI
----------------------------------------------
(Intercept)        |       0.00 | [0.00, 0.00]
daysSinceFirstText |       0.07 | [0.06, 0.07]
# Standardization method: pseudo

Parameter          | Std. Coef. |       95% CI
----------------------------------------------
(Intercept)        |       0.00 | [0.00, 0.00]
daysSinceFirstText |       0.03 | [0.02, 0.04]

Differences in Symptoms Across Clusters

Cluster predicting fdSx controlling for bSx

Plot

Do clusters moderate?

Table

Plot

All plots together

Q5) Is this occurring for all words, or is “I” unique in some way?

Changes over Therapy

Track with Changes in Symptoms